[SPARK-55714][SQL] JDK might throw ArithmeticException without message

[pan3793]

What changes were proposed in this pull request?

Canonicalize the error message of SparkArithmeticException:

• null => "overflow"
• "byte overflow", "long overflow", etc., to "overflow".
• leave others as-is.

Update tests to accept null message from ArithmeticException

Why are the changes needed?

During test, I found JDK 25 might throw ArithmeticException with null message, and it varies on different platforms / CPU models, for example, it's different on Apple M1 and x86 Server, and even different on Intel Skylake and AMD Zen4 (both are x86 but with different instructions support)

https://bugs.openjdk.org/browse/JDK-8367990

This is expected behavior. For "hot throws", the JIT will produce compiled code that does not go through the interpreter via deoptimization to throw an exception but throws a pre-allocated exception object without a stack trace or message. ... this optimization can be disabled with -XX:-OmitStackTraceInFastThrow.

Does this PR introduce any user-facing change?

Some ARITHMETIC_OVERFLOW error condition message changes from "byte overflow", "long overflow", etc., to "overflow".

How was this patch tested?

Pass GHA for JDK 17. JDK 25 will be covered by daily tests.

Was this patch authored or co-authored using generative AI tooling?

No, the patch is crafted by hand.

[dongjoon-hyun]

Can we use a simple Scala-style one-liner?

-    val canonicalMessage = message match {
-      // For "hot throws", the JIT will produce compiled code that does not
-      // go through the interpreter via deoptimization to throw an exception
-      // but throws a pre-allocated exception object without a stack trace
-      // or message. See https://bugs.openjdk.org/browse/JDK-8367990
-      case null => "overflow"
-      case m if m.contains("overflow") => "overflow"
-      case m => m
-    }
+    val canonicalMessage = Option(message).filterNot(_.contains("overflow")).getOrElse("overflow")